LAUSR

Introduction Epilepsy patients are often susceptible to cardiorespiratory adverse events during and immediately after the seizure. Unwitnessed seizures during sleep are associated with the highest rate of SUDEP - a common cause of mortality in epilepsy. To improve surveillance, different wearable sensors (like seizure warning watches) are developed for seizure advisory systems. These devices are restricted to detecting ictal events from heart rate variability, motion, and muscle activities. As the risk of cardiorespiratory complications extends beyond seizure to postictal period, it may be worthwhile to develop novel extracranial physiological biomarkers that can be used to estimate the severity of seizures from the postictal period. During a convulsive seizure, movement artifact often precludes reliable data interpretation, and hence the motivation of this study is to identify parameters at the post ictal period when artifact is expected to be less. Using advanced signal processing and machine learning tools we systematically analyzed multiple parameters obtained from a single-channel electrocardiogram to estimate the severity of seizure and epilepsy. Methods Patients admitted to a level-IV epilepsy center for video scalp EEG monitoring were included in this study. Only patients with confirmed seizures were selected. A single lead EKG was recorded simultaneously with video EEG (Natus Xltek system). Sampling frequency was 256 Hz. The EKG was analyzed in time-frequency and nonlinear domains - mean HR, standard HR, mean RR interval, RMSSD, HF, LF, approximate and sample entropy. Two minutes of epoch were selected for a baseline (at least 6 h apart from a seizure), pre seizure and post seizure. The study goal was to identify parameters following seizure termination hence seizure was not analyzed. ANOVA was performed to select features and machine learning tools (deep learning and SVM) were used for classification. Seizure severity was scored by using duration, type (focal vs. generalized seizure), Chalfont Scale and SUDEP score. Results Forty patients (M = 28) with a mean age of 32-years were included in this retrospective study. Seventy-eight seizures (49 focal seizures) were analyzed. In the postictal period, the mean and standard deviation of HR increased while the mean RR decreased significantly in comparison to baseline or pre-seizure state. Both sample and approximate entropy decreased significantly in the postictal state. Deep learning ML classified post ictal state from pre ictal or baseline with higher accuracy (>90%) by using combination of approximate entropy, mean RR. However, other features like LF, HF are still analyzed for higher accuracy and to estimate seizure severity. Conclusion In this novel study, we aim to identify physiological parameters that can estimate severity of seizures from post ictal EKG recordings. Our preliminary analyses shows post ictal mean RR, HR and approximate entropy as candidate reliable markers for estimating seizure severity.